Successfully reported this slideshow.
Your SlideShare is downloading. ×


Loading in …3

Check these out next

1 of 16 Ad

More Related Content

Similar to athena-neuroinformatics-2018 (20)

Recently uploaded (20)



  1. 1. Automating annotations of the cognitive neuroimaging literature using ATHENA Riedel MC, Salo T, Hays J, Turner MD, Sutherland MT, Turner JA, & Laird AR Neuroinformatics and Brain Connectivity Lab
  2. 2. Neuroimaging Research • Increasing in volume and scope • Embedded in this literature is knowledge capturing a system-level probing of functional brain organization • The challenge for cognitive neuroscience is harnessing this knowledge and translating it into improved neurocognitive models 0 2000 4000 6000 8000 10000 12000 14000 2000 2002 2004 2006 2008 2010 2012 2014 Published Neuroimaging Studies
  3. 3. Cognitive Paradigm Ontology • Knowledge modeling effort to study the relationship between brain structure and function • Seeks to represent stimuli, responses, and instructions that define conditions of an fMRI experiment in a standardized format • System of labels for annotating neuroimaging articles
  4. 4. Cognitive Paradigm Ontology Behavioral Domain Paradigm Class Diagnosis Instruction Context Stimulus Modality Stimulus Type Response Modality Response Type Action Cognition Emotion Interoception Perception Anger Fear Happiness Sadness n-back Face Monitor/Discrimination Classical conditioning Delay discounting Film viewing Go/No-Go Autism Spectrum Disorders Bipolar Disorders Depression Normal Schizophrenia Attend Count Detect Discriminate Recall Disease Effects Drug Effects Normal Mapping Auditory Tactile Visual Digits Faces Letters Pictures Shapes Hand None Oral/Facial Button Press None Speech
  5. 5. Goals • Develop framework for automated annotations of neuroimaging articles • Evaluate classifier performance across variable parameters: • corpus • feature space • classification algorithm • Characterize relationships between labels by assessing similar vocabularies used for classification Problem • Manual annotation is time-limiting, field is too large • Bias/human error
  6. 6. Classification Features • Property or characteristic of something being measured • Related to explanator variables in linear regression • Examples: • Speech recognition: noise ratios, length of sounds, relative power, filter matches • Spam detection: email headers, email structure, language, term frequency • Character recognition: histogram counts of black pixels in horizontal and vertical direction, number of internal holes, stroke detection
  7. 7. Background-Studies incorporating direct comparisons across all phases of bipolar (BP) disorder are needed to elucidate the pathophysiology of bipolar disorder. However functional, neuroimaging studies that differentiate bipolar mood states from each other and from healthy subjects are few and have yielded inconsistent findings. Feature Spaces bag-of-words Cognitive Atlas bipolar bipolar disorder disorder bipolar bipolar mood bipolar mood states mood states mood states bipolar disorder mood
  8. 8. Classification Procedure neuroimaging article n = 2,633 Behavioral Domain Context Diagnosis Instruction Paradigm Class Response Modality Response Type Stimulus Modality Stimulus Type abstracts-only full-text CogPO Labels corpora text extraction bag-of-words Cognitive Atlas feature spaces training/test dataset splits k = 5 80%/20% feature vectorization and reduction f = 1,754 parameter tuning k = 2 classification Bernoulli naïve Bayes k-nearest neighbors logistic regression support vector classifier cross-validation 100 iterations
  9. 9. Assessing Classifier Performance • Classifier performance evaluated using F1-score • 𝐹1 = 2 × 𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛×𝑟𝑒𝑐𝑎𝑙𝑙 𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛+𝑟𝑒𝑐𝑎𝑙𝑙 , 𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = 𝑡𝑝 𝑡𝑝+𝑓𝑝 , 𝑟𝑒𝑐𝑎𝑙𝑙 = 𝑡𝑝 𝑡𝑝+𝑓𝑛 • Ranges from 0 to 1 • F1-scores averaged across labels for overall performance
  10. 10. Classifier Performance F1-score
  11. 11. Representation of Classification Features • bag-of-words features used to classify each label for Behavioral Domain and Paradigm Class • Used distributions of feature representation to calculate correlation matrix • Regressed co-occurrence of labels from correlation coefficients • Performed hierarchical clustering on resulting matrix to assess similarity of classification features between labels
  12. 12. ParadigmClass_Reward BehavioralDomain_Cognition_SocialCognition BehavioralDomain_Cognition_Memory_Explicit BehavioralDomain_Cognition_Attention BehavioralDomain_Perception_Vision_Shape BehavioralDomain_Perception_Vision BehavioralDomain_Perception_Audition ParadigmClass_WordGeneration BehavioralDomain_Cognition_Language_Speech BehavioralDomain_Cognition_Language_Semantics BehavioralDomain_Cognition_Language ParadigmClass_Reading BehavioralDomain_Action_Execution_Speech BehavioralDomain_Perception ParadigmClass_Stroop ParadigmClass_GoNoGo BehavioralDomain_Action_Inhibition ParadigmClass_EmotionInduction BehavioralDomain_Emotion_Happiness ParadigmClass_FaceMonitorDiscrimination BehavioralDomain_Emotion_Fear BehavioralDomain_Emotion ParadigmClass_SemanticMonitorDiscrimination BehavioralDomain_Cognition_Memory_Working ParadigmClass_PassiveViewing ParadigmClass_nback ParadigmClass_DelayedMatchtoSample BehavioralDomain_Cognition_Reasoning ParadigmClass_Encoding ParadigmClass_CuedExplicitRecognitionRecall BehavioralDomain_Cognition_Memory ParadigmClass_FingerTappingButtonPress ParadigmClass_VisuospatialAttention BehavioralDomain_Action_Execution BehavioralDomain_Action BehavioralDomain_Interoception BehavioralDomain_Cognition BehavioralDomain_Perception_Vision_Motion ParadigmClass_Rest BehavioralDomain_Action_Rest BehavioralDomain_Perception_Somesthesis ParadigmClass_PainMonitorDiscrimination BehavioralDomain_Perception_Somesthesis_Pain LanguageEmotionMemoryPain
  14. 14. Conclusions and Future Works • full-text, bag-of-words performed best • Cognitive Atlas features outperform bag-of-words when only using text from abstracts • Anatomical terms dominate features for classification when using bag-of-words • Test on independent dataset • Validate by replicating existing meta-analyses • Specify Cognitive Atlas • Integrate with existing frameworks
  15. 15. Acknowledgements External Collaborators Dr. Angela Laird Dr. Matthew Sutherland Dr. Michael Tobia Dr. Veronica Del Prete Jessica Bartley Katherine Bottenhorn Jessica Flannery Ranjita Poudel Taylor Salo Lauren Hill Chelsea Greaves Rosario Pintos Lobo Laura Ucros Diamela Arencibia Jennifer Foreman Ariel Gonzalez Neuroinformatics and Brain Connectivity Lab Jessica Turner Matthew Turner Neuroinformatics and Brain Connectivity Lab NSF 1631325 NSF REAL DRL1420627 NSF CNS 1532061 NIH R01 DA041353 NIH U01 DA041156 NIH K01 DA037819 NIH U54 MD012393
  16. 16. Classifiers • Bernoulli naïve Bayes • Trains on binary word occurrence vectors instead of word counts • logistic regression • Linear model for classification • k-nearest neighbors • Identifies nearest k articles in distance and uses majority vote to determine if its about a label • support vector machine • Creates high-dimensional decision hyper-plane

Editor's Notes

  • Thank you for the opportunity to speak here today. Im going to present some work we have been doing using classification techniques to automatically annotate the neuroimaging literature with labels from a cognitive ontology.
  • Neuroimaging research has seen an explosion of growth over the past 20 years, indicated here by the number of published neuroimaging studies per year. As we have continued to explore the relationship between brain structure and function, the research has also become increasingly complex.
    With such a wealth of information embedded in this literature about brain organization, the challenge for us as neuroscientists is to harness this knowledge in a digestible manner in a way that enhances our understanding of neurocognitive models.
  • The Cognitive Paradigm Ontology or CogPO is a step in this direction, providing a discrete set of terms meant to help study the relationship between brain structure and function.
    These terms are meant to characterize elements of experimental design like stimuli, responses, instructions, and can be used to annotate neuroimaging articles to provide concise references about the research in that that article.
  • Briefly, CogPO consists of 9 dimensions. One of which is behavioral domain, which describes a mental construct and contains labels such as Action, Cognition, Emotion, Interoception and Perception. And these labels may be even more descriptive such as Anger, Fear, or Happiness for Emotion.
    Then Paradigm Class describes different tasks, such as n-back, delay-discounting and go/no-go. Then there are other types of labels such as Diagnosis, Instruction, Context, and Response and Stimulus Modality and Type.
    Thus, CogPO informs cognitive models by being able to synthesize articles with similar labels to perform meta-analyses.
  • The problem we are currently facing is that with such a large amount of research available, manual annotation of the literature is time-limiting and nearly impossible to keep up with. Plus add in bias and human error associated with manual annotations.
    Therefore, we sought to develop a framework for annotating neuroimaging articles in an automated manner using the CogPO labels.
    To do this, we wanted to evaluate classification performance by varying three parameters: corpus, features, and different classification algorithms.
    Then we wanted to characterize the relationships between labels by assessing the most frequently used features in the classification process. That is, can we use data-driven approaches to determine if neuroimagers are using similar vocabularies to describe certain cognitive paradigms.
  • First, when I talk about classification features, they are a measurable property or characteristic that can be used for classification.
    They are similar to the explanatory variables in a linear regression.
    Some examples of features in classification are noise ratios and lengths of sound in speech recognition, and email text or headers in spam detection.
  • We wanted to evaluate the performance of two types of features: terms extracted from the text, which is the bag-of-words approach, and representation of terms defined by the Cognitive Atlas. A description of the Cognitive Atlas deserves more time here, but briefly, it is a vocabulary of about 1700 terms describing concepts, tasks, disorders, and theories in cognitive science. What also makes the Cognitive Atlas unique is the relationships between terms, such as working memory is a KIND OF memory, and a PART OF decision making.
    Here is a short example of how the bag of words and Cognitive Atlas approaches for defining features differs. Consider this small text.
    The bag-of-words approach would take the words “bipolar disorder” and break it into “bipolar” “bipolar disorder” and “disorder”, and “bipolar mood states” can then be broken into all combinations of three or less terms shown here. I should mention that all terms in this text could be used for classification, Im just focusing on these two example for illustrative purposes.
    Then, in the Cognitive Atlas approach, only terms defined by the Cognitive Atlas are used, so the only terms that would be used for classification would be “bipolar disorder” and “mood”, and all other terms in this text are ignored.
  • Now I’ll walk you through how we defined our classifiers for each CogPO label.
    We utilized a dataset of 2,633 neuroimaging articles that were manually annotated with CogPO labels.
    As I mentioned before, we evaluate extracting features from either just the abstracts or the full-text.
    Once text was extracted according to the bag-of-words or Cognitive Atlas feature space, we performed 100 iterations of a repeated 5-fold cross-validation procedure. In this procedure, in each iteration, the dataset was split into 5 folds, where each fold was divided into a training dataset, which consisted of 80% of the articles, and a test dataset, which consisted of 20% of the articles.
    We then vectorized the features based on frequency of appearance in an article and incorporated the frequency of that feature across all articles. Because the Cognitive Atlas only consisted of 1,754 terms, we reduced the bag-of-words terms using a chi-square test that removes all but the top 1,754 features for a particular label.
    Then, in preparation for classification, we performed a 2-fold cross-validation procedure to tune the hyperparameters for classification. Depending on the classification algorithm, this step basically optimizes the cost function and smoothing kernel.
    Finally, we used 4 different classification algorithms to generate a classifier for each CogPO label, logistic regression, Bernoulli naïve bayes, k-nearest neighbors, and support vector machine.
  • To assess classifier performance, we used F1-scores, which are dependent on precision and recall.
    The F1-score can range from 0 to 1, where 1 represents perfect classification
    We calculated the F1-score for each label and averaged across all 100 iterations. Then, to determine which combination of corpus, feature space, and classification algorithm performed best, we averaged across all CogPO labels.
  • This graph represents overall performance, separated by classification algorithm on the bottom. Abstracts are in blue, full-text is in orange, bag-of-words are represented by circles and Cognitive Atlas by X’s.
    Here, the top performer used full-text, bag-of-words, and the logistic regression algorithm.
    Its also worth noting that when using the support vector machine algorithm, the cognitive Atlas approach did not differ that greatly from the bag-of-words approach when using full-text.
    And perhaps more interesting, the Cognitive Atlas feature space approach actually outperformed the bag-of-words approach when only using article abstracts. This could be particularly useful for two reasons: 1) Cognitive Atlas provides a platform for classifying based on an ontology specifically designed for the cognitive sciences, and 2) currently abstract-text are more accessible than full-article text, and may provide a means for annotating a larger proportion of the literature.
  • Since the bag-of-words approach performed the best, we wanted to determine which CogPO labels in Behavioral Domain and Paradigm Class used the same features for classification across iterations. This provides insight into vocabularies used to discuss similar constructs.
    We used the distribution of feature representations to calculate a correlation matrix, and corrected for the fact that some labels tend to be assigned together a lot, such as Emotion and Emotion.Fear.
    Then we performed hierarchical clustering on the resulting matrix to provide a visual representation of similar labels.
  • Here is the dendrogram, and just based on visual inspection of the dendrogram, we isolated four clusters of labels. You can see that each cluster contains labels from both Paradigm Class and Behavioral Domains.
    The green cluster contains labels related to cognition, perception and language. We created a word cloud of the top 10% of features within the labels associated with this cluster and see dominant terms such as temporal gyrus, anterior cingulate. This cluster seems to be dominated by anatomical terms.
    The blue cluster seems primarily related to inhibition, and the resulting word cloud exhibits cingulate cortex, anterior cingulate, “event related”. Now we can see some terms related to task design involved in the classification process.
    The purple cluster is pretty large, containing terms related to emotion and memory. The resulting word cloud exhibits terms like working memory, prefrontal cortex, facial expression, and major depressive disorder. Here disorders frequently studied within a domain become prominent in addition to more information about task design.
    Finally, the red cluster contains terms related to pain and action. The resulting word cloud contains terms such as reaction time, working memory. This may be a little less informative which isn’t that surprising given the diversity of the labels assigned to the cluster.
    We can also see tight groupings of labels related to specific constructs, such as language, emotion, memory, and pain.
  • While these topics are somewhat subjectively chosen, we generated word clouds for each one.
    Within the language topic we can see terms such as superior temporal being dominant, and amygdala, emotion, and fusiform dominant for emotion. The memory topic contains terms like working memory, prefrontal cortex, dorsolateral, and cingulate. Finally the pain topic contains terms such as anterior cingulate and insula.
    Again, we can see here that related to specific constructs, anatomical terms really seemed to dominate the features used for classification. But it does demonstrate that within constructs, neuroimagers are discussing similar brain structures!
  • To wrap everything up, we evaluated classifier performance for CogPO labels using text from either abstracts or the full article, bag-of-words features or Cognitive Atlas terms, and different classification algorithms. We found that the combination of full-text, bag-of-words, and logistic regression performed the best.
    The Cognitive Atlas features outperformed the bag-of-words features when only using text from the abstracts.
    Anatomical terms dominated the features used for classification when using bag-of-words.
    Our future works include testing on an independent dataset, and validating these classifiers by replicating existing meta-analyses of manually annotated articles.
    We would additionally like to assist in the process of fully specifying the relationships between terms in the Cognitive Atlas and seek to integrate these classifiers in existing frameworks.
  • I would like to thank everyone in the Neuroinformatics and Brain Connectivity Lab for their contributions to this project, especially Taylor Salo. Id also like to thank our collaborators at Georgia State for project and analysis development. And of course thank you again for inviting me to present our work here today.
    I’ll take any questions you have at this time.
  • Logistic regression – probabilities describing the possible outcomes of a single trial are modeled using the S-shaped logistic function